Rapid classification and quality assessment of citrus essential oils via machine-learning-assisted Raman spectroscopy

Abstract

Citrus essential oils (EOs) require accurate identification and assessment to ensure authenticity and consistency. However, conventional techniques such as gas chromatography (GC) and mass spectrometry (MS) are time-consuming and expensive, highlighting the need for novel analytical methods. This study proposes an approach for EOs detection using Raman spectroscopy (RS) combined with machine learning (ML) algorithms. Six citrus EOs underwent an evaporation experiment, with Raman spectra collected at five time points and GC-MS was used to analyze compositional changes at the starting and ending points of evaporation as a standard reference. Five ML algorithms were developed to identify differences among EOs and monitor their temporal changes. The predictive performance was evaluated using multiple quantitative metrics. The results show that the support vector machine (SVM) consistently achieves the best performance across all prediction tasks, and the interpretability algorithm identified key components in the Raman spectra of EOs. Taken together, RS-SVM is proven to be an accurate and cost-effective analytical technique for the rapid identification and quality assessment of citrus EOs.

Graphical abstract

Highlights

• •The volatilization rates of six citrus EOs were measured over twenty-eight days.
• •Components of six citrus EOs were characterized via GC-MS as a reference.
• •The Raman spectra of six citrus EOs were collected, visualized, and characterized.
• •Machine models were built for the identification and assessment of six citrus EOs.
• •Model interpretability found key components in the Raman spectra of six citrus EOs.

1. Introduction

Citrus essential oils (EOs) are natural products extracted from the peel of citrus species (Felicia et al., 2024) and are widely used in the food, pharmaceutical, and cosmetic industries due to their unique aromatic properties and biological activities (Bora et al., 2020). A citrus EO is a complex mixture composed of both volatile and non-volatile components. The volatile organic compounds (VOCs) in citrus EOs include monoterpenes, sesquiterpenes, and their oxygenated derivatives, with limonene being the major constituent (Agarwal et al., 2022; Han et al., 2024). Due to their chemical nature, the VOCs in EOs are prone to oxidation and polymerization under environmental factors such as temperature, light, and oxygen, leading to a reduction in their quality and pharmacological properties (Turek and Stintzing, 2012, 2013). Such instability not only affects their therapeutic efficacy but also poses challenges for quality assurance and market regulation. To ensure consumer safety and fair trade, the quality of EOs must be assessed using analytical techniques, including physical, organoleptic, chemical, chromatographic, and spectroscopic methods (Do et al., 2015).

Conventional sensory and physicochemical tests are informative but insufficient for complex EO matrices. Advanced analytical techniques, including chromatographic and spectroscopic methods, are therefore required for more accurate and detailed analysis. Chromatographic techniques, such as high-performance liquid chromatography (HPLC), liquid chromatography-mass spectrometry (LC-MS), and gas chromatography-mass spectrometry (GC-MS), are used to separate compounds in EOs and to perform both qualitative and quantitative analyses. Meanwhile, spectroscopic techniques, such as ultraviolet–visible spectroscopy (UV–Vis), infrared spectroscopy (IR), nuclear magnetic resonance spectroscopy (NMR), and Raman spectroscopy (RS), can be applied to identify the molecular structures of EO components (Ali et al., 2023; Tang et al., 2025). Among these analytical methods, GC-MS has become the gold standard for identifying VOCs in EOs and verifying their authenticity, owing to its high sensitivity and accuracy (Ali et al., 2023; Ratnasekhar et al., 2025). It provides detailed chromatograms to identify components, but its limitations, such as high cost, complex operation, and long analysis time, restrict its application in on-site quality detection (Do et al., 2015; Epping and Koch, 2023).

In recent years, researchers have begun to explore novel detection technologies, such as RS, combined with chemometrics, to achieve rapid identification and quality evaluation of EO samples (Cebi et al., 2021; Gloerfelt-Tarp et al., 2022; Lebanov and Paull, 2021; Nelson et al., 2020). RS provides detailed fingerprint like information (Huang et al., 2022) and offers high sensitivity, rapid analysis, and good accuracy, making it well suited for on-site applications (Cebi et al., 2021). However, distinguishing highly similar citrus EO spectra remains challenging. The application of machine learning (ML) algorithms provides a new approach to address this issue, as they can leverage the powerful pattern recognition capabilities to reveal subtle differences in Raman spectra. Recent studies have shown that RS combined with ML has emerged as a powerful tool for the identification, classification, and authenticity testing of EOs (Huang et al., 2022; Wen et al., 2025). For example, a smartphone-based handheld Raman spectrometer was used to detect and quantify adulteration in EOs through the ML analysis of Raman spectra (Lebanov and Paull, 2021). However, existing studies have mainly focused on detecting adulteration in EOs, with few addressing the challenge of assessing how their composition and quality change under conditions that simulate real-world use, such as evaporation. We developed a novel approach that not only identifies six citrus EOs with high similarities but also monitors Raman spectral changes resulting from air-exposed evaporation and deterioration. This work extends the application of RS-ML methods beyond static classification, offering a dynamic and interpretable framework for citrus EO authentication and quality assessment.

In this study, six citrus EOs (buddha's hand, grapefruit, lemon, lime, mandarin orange, and sweet orange) were subjected to a 28-day evaporation experiment. Raman spectra were collected at five time points (0, 7, 14, 21, and 28 days). GC-MS analysis identified the VOCs in six citrus EOs, indicating both EO authenticity and compositional differences in the VOCs before and after evaporation. Average and deconvoluted spectra were used to identify characteristic peaks. Clustering analysis revealed intrinsic differences among the six EOs, as well as among the five evaporation time points within each EO. Multiple ML algorithms were developed to classify the six EOs and to monitor evaporation stage changes over time, which serves as the basis for quality assessment in this work. Accuracy, precision, recall, and F1-score were used to evaluate model performance on an independent external validation set. The results show that the support vector machine (SVM) can accurately identify and track evaporation-related changes in citrus EOs. Interpretability analysis revealed the contribution of different features to the model's decisions, supporting the reliability of SVM classification. In summary, this study developed a novel RS-SVM-based technique for the rapid and accurate identification and quality evaluation of citrus EOs, addressing the current limitations of traditional GC-MS analysis and holding potential for real-world applications.

2. Materials and methods

2.1. Sample preparation

Grapefruit EO (CAS 8016-20-4), Mandarin orange EO (CAS 8008-31-9), Sweet orange EO (CAS 8008-57-9), Buddha's hand EO (CAS 8007-75-8), and Lemon EO (CAS 8008-56-8) were purchased from Shanghai Macklin Biochemical Co., Ltd. (China). Lime EO (CAS 8008-26-2) was bought from Shanghai Chemisci Chemical Technology Co., Ltd. (China).

2.2. Evaporation experiments

In this study, 25 mL of each EO was placed in 30 mL glass vials (Bikeman Bio, China), with three biological replicates per EO (n = 18). The vials were stored in the dark for a 28-day evaporation experiment. Barometric pressure was monitored using a MIAOXIN TH20R barometer (Pingyang Miaoguan Technology Co., Ltd., Zhejiang, China), while temperature and humidity were recorded with a Wissdom system (Beijing Wisdon Network Technology Co., Ltd., Beijing, China). Sample weights were recorded daily. Raman spectra were collected at five time points (0, 7, 14, 21, and 28 days). Additionally, 10 mL each of pure EO at the starting point and after 28 days of evaporation were prepared for GC-MS analysis.

2.3. GC-MS analyses

The GC-MS analysis method from Cebi et al. was adopted and modified to analyze six citrus EOs before and after a 28-day evaporation experiment (Cebi et al., 2021). Analysis was performed by using Agilent 7890A-5975C (Agilent Technologies, USA), with a DANI HSS86.50 headspace sampler from DANI, Italy. 20 mL headspace vials were used to transfer all samples, which were sealed and heated at 120 °C for 20 min. The injection valve temperature was set to 150 °C, and the transfer line temperature was set to 180 °C. A sample volume of 1 mL was injected. The chromatographic separation was carried out on a DB-WAX column (30.0 m $\times$ 250 μm, 0.25 μm), with the following temperature program: the initial temperature was 50 °C, held for 5 min, then increased at a rate of 5 °C/min to 150 °C, held for 5 min, and then increased at 10 °C/min to 230 °C, held for 5 min. The injector temperature was 250 °C, the transfer line temperature was 240 °C and helium was used as the carrier gas at a flow rate of 1.0 mL/min, with a split ratio of 100:1. The mass spectrometer operated in Electron Ionization (EI) mode, with an electron energy of 70 eV, an ion source temperature of 230 °C, and a quadrupole temperature of 150 °C. The scan mode was set to “Scan” with a mass range of 20–500 u. The identification of volatile compounds was performed by comparing the total ion chromatograms with the commercial mass spectra libraries (NIST20). Since no internal standard was used, the GC-MS results were interpreted qualitatively rather than quantitatively. Only the top ten compounds with the highest relative contents were considered, accounting for more than 90 % of the total VOC composition, while trace compounds were excluded to avoid interference from impurities (Supplementary Tables S1–S2). Compounds were grouped by functional class, and the relative contribution of each class was calculated and visualized using bar charts.

2.4. Collection of Raman signals

Raman signals were detected using Anton Paar™ Cora100 Handheld Raman spectrometer (Anton Paar GmbH, Austria), configured with the following parameters: (i) excitation wavelength: 785 nm, (ii) excitation power: 300 mW, (iii) spectral resolution: 10 cm⁻¹, (iv) spectrum acquisition time: 8 s, and (v) detection spectral range: 400-2300 cm⁻¹. The desktop software Cora 100 Connect (Anton Paar GmbH, Austria) was used to generate Raman spectra, manage spectral databases, and output spectral data. A total of 18 samples were analyzed, and for each sample, 30 spectra were collected from three different measurement angles during collection to ensure data stability and reproducibility.

2.5. Average and deconvoluted Raman spectra

To analyze the Raman spectral characteristic peaks of citrus EOs at five time points, the average signal intensity of all Raman signals at each Raman shift point for each sample was calculated to generate a representative average Raman spectrum. Origin Software (OriginLab, United States) was used to visualize a shaded area representing the standard deviation (SD) around the average Raman spectrum, providing an intuitive depiction of the distribution and variability of the signal intensity. The software's fitting peak function Pro was used to automatically fit Raman spectral feature peaks, facilitating the identification of molecular vibrations and components. Spectral deconvolution was also applied to the average Raman spectra of each sample to reveal subtle spectral differences between different types of citrus EOs and across various evaporation times. The Voigt function in Origin software, which represents the convolution of Lorentzian and Gaussian widths, was used for spectral deconvolution. All Lorentzian and Gaussian width values for the feature peaks were set to 1, and the fitting process continued until convergence, providing an accurate description of the shape and width of each characteristic peak.

2.6. Cluster analysis

OPLS-DA analyses were performed on the Raman spectra obtained from citrus EO samples (refer to the RS measurements section), using SIMCA 14.1 (Umetrics, Umea, Sweden)(Yao et al., 2024). For the citrus EO classification model, Raman spectra obtained before and after 28 days of evaporation were analyzed. Citrus type served as the dependent variable (Y), with Raman spectra as the independent variable (X). For the citrus EO quality detection model, Raman spectra measured at five time points for each citrus EO were analyzed. The five time points served as the dependent variable (Y), with Raman spectra as the independent variable (X). OPLS-DA model extracts meaningful information from the spectra to differentiate between various citrus types or evaporation days. The X and Y variables were used to build the OPLS-DA model, with cross-validation employed to calculate R2X, R2Y, and Q² values. R2X indicates how well the model explains variance in the independent variable (Raman spectral). R2Y reflects the variance in the dependent variable (citrus type or time points) explained by the model. Q² measures the model's ability to predict new, unseen data. Higher value indicates better generalization and predictive power (Wang et al., 2025).

2.7. ML algorithms and interpretability analysis

Prior to inputting the data into ML algorithms, the dataset was split using the train_test_split function, with 80 % allocated to the training set and 20 % to the external test set. The training set was further split in an 8:2 ratio into an internal training set and an internal validation set, which were used for model optimization, training, and overfitting assessment. The external test set was reserved exclusively for evaluating model performance on unseen data. Five ML algorithms implemented in the scikit-learn data analysis library (version 1.3.0), including Adaptive Boosting (AdaBoost), Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), and eXtreme Gradient Boosting (XGBoost), were employed to train on the Raman spectral data. Hyperparameter optimization for all models was conducted using GridSearchCV (Supplementary Tables S3–10), and the optimal hyperparameter configurations identified were subsequently used for model training. Model performance was evaluated using Accuracy, Precision, Recall, F1-score and 5-fold cross-validation (5-Fold CV) metrics, while the generalization capability of each model was assessed through 5-fold cross-validation. Interpretability analysis was performed based on the class-specific weight vectors coef. By averaging these weight vectors across all classes, the mean contribution of each feature was computed, reflecting its overall importance in the classification process. These contributions were visualized as scatter plots to elucidate the role of different Raman shifts in the model's decision-making process.

2.8. Statistical analysis

To the analysis of evaporation rate differences and Raman spectral differences across five time points, the Shapiro-Wilk test was performed to assess the normality of the data. If the P-value is greater than 0.05, the data were considered normally distributed, and parametric tests were applied. Otherwise, nonparametric tests were used. One-way ANOVA was conducted to compare the average evaporation rates across three parallel experiments for each EO, followed by Tukey's post hoc test. The Kruskal-Wallis test was used to compare the average peak intensities across three replicates at five time points for each EO, followed by Dunn's post hoc test. The significance level was set at a P-value of less than 0.05 for all tests. Differences between groups are denoted by letters. Groups sharing the same letter are not significantly different, whereas groups with different letters exhibit significant differences. The results were analyzed using GraphPad Prism software. Evaporation rate differences are presented as bar graphs, and spectral differences at different time points are depicted using violin plots.

3. Result

3.1. Evaporation and composition of citrus EOs

A 28-day evaporation experiment was conducted on six citrus EOs, with continuous monitoring of room temperature, atmospheric pressure, relative air humidity, and the weights of the EOs. The average temperature was 23.43 ± 1.07 °C, atmospheric pressure was 101.02 ± 0.27 hPa, and relative air humidity was 57.02 ± 11.15 % (Fig. 1a). The stable environmental conditions enabled an accurate assessment of the evaporation behavior of citrus EOs. The average weight change of the six citrus EOs over the 28 days is shown in Fig. 1b, with grey shading indicating the variation among three independent replicate samples. The corresponding weight changes are provided in Supplementary Table S11 and S12. The evaporation rates of six citrus EOs were depicted in a bar chart as shown in Fig. 1c. The bars represent the mean evaporation rate of three replicate samples, with error bars indicating the standard error of the difference. Alphabetical annotations denote statistical differences among groups. Groups sharing the same letter(s) are statistically indifferent, while groups with different letters are significantly different. Specific values and calculation methods were present in Supplementary Table S13. The citrus EOs are ranked by mean evaporation rates from high to low values as follows: lemon, buddha's hand, sweet orange, lime, mandarin orange, and grapefruit. Based on these rates, the citrus EOs were grouped into three categories: grapefruit, with the slowest rate, and lemon, buddha's hand, and sweet orange, with faster rates, and lime and mandarin orange, with moderate rates. Significant differences were observed among the groups, but no significant differences were found within each group. The variability in evaporation rates is attributed to the distinct chemical compositions of the citrus EOs. These differences in evaporation behavior are likely associated with the compositional characteristics and chemical reactivity of the volatile constituents.

Fig. 1.

Volatilization rates and relative content changes in the top 10 VOCs of six citrus EOs over 28 days. (a) Environmental parameters during volatilization: temperature (23.43 ± 1.07 °C), atmospheric pressure (101.02 ± 0.27 hPa), humidity (57.02 ± 11.15 %). (b) Weight change during volatilization; (c) Statistical differences in volatilization rate; (d) VOC categories based on functional groups. For statistical analysis, Tukey's test was conducted. The results were shown via alphabetical annotations that denote statistical differences among groups. Groups sharing the same letter(s) are statistically indifferent, while groups with different letters are significantly different.

Citrus EOs are composed of numerous highly volatile compounds from various chemical classes (Turek and Stintzing, 2013). Therefore, GC-MS analysis was conducted on six citrus EOs before and after evaporation to characterize the variation in VOCs. The identified VOCs were grouped into six functional classes, namely terpenes, alcohols, ketones, oxides, aromatic compounds, and aldehydes. Similar VOC classes were observed across all six citrus EOs. Among these, terpenes were the predominant constituents (Agarwal et al., 2022). D-Limonene was identified as the major compound, accounting for approximately 63–86 % of the total VOC content, while alcohols, ketones, oxides, aromatic compounds, and aldehydes were present only in trace amounts. The VOC profiles of Buddha's hand EO (Yang et al., 2025), grapefruit EO (Ahmed et al., 2019), lemon EO (Javed et al., 2014), mandarin orange EO (Javed et al., 2014), lime EO (Chisholm et al., 2003), and sweet orange EO (Qiao et al., 2008) were consistent with previously documented findings. These findings support the authenticity and reveal compositional differences among the citrus EOs. To investigate the temporal variation in chemical composition, we further compared the relative contents of the top 10 VOCs in each citrus EO before and after volatilization. These 10 compounds accounted for more than 90 % of the total VOC content. As shown in Fig. 1d, the relative contents of individual compounds varied in all six EOs, indicating a significant alteration in the released VOC profiles after 28 days of volatilization. Except for grapefruit EO, the relative contents of terpenes decreased in all samples, with the most pronounced decline observed in mandarin orange EO, while the proportion of alcohols increased across all six citrus EOs.

3.2. Raman spectral analysis and clustering of citrus EOs

The average Raman spectra (solid lines) of citrus EOs were collected for both the raw EOs and those after 28 days of evaporation, with grey areas indicating the standard deviation of replicate sample measurements (Fig. 2a & b). To evaluate the reproducibility of Raman spectral measurements, variations across different sampling points were analyzed by calculating the relative standard deviation (RSD) of four prominent characteristic peaks from 20 spectra for each citrus EO (Liu et al., 2025). The RSD values for all tested EOs ranged from 0.81 % to 2.65 % (Supplementary Fig. S1). Specifically, for Buddha's hand EO (Supplementary Fig. S1a), the four prominent peaks are located at 758 cm⁻¹ (grey), 1436 cm⁻¹ (red), 1644 cm⁻¹ (yellow), and 1678 cm⁻¹ (green). The RSDs of the different peaks are 1.12 %, 1.25 %, 1.50 %, and 1.55 %, all indicating low variability, with similar reproducibility observed across the other five citrus EOs. These consistently low RSD values (<3 %) confirm that the Raman spectra measured in this study exhibit good repeatability (Mou et al., 2024). Spectral deconvolution was applied to resolve overlapping signals within the spectra. The deconvoluted bands show that the characteristic peaks of each EO are clearly separated (Fig. 2c & d). An OPLS-DA model was then performed for supervised classification of citrus EOs (Fig. 2e), revealing six distinct clusters corresponding to the six citrus species. The model demonstrated strong generalizability and predictive power, with R2X = 0.988, R2Y = 0.93, and Q2 = 0.926. After 28 days of evaporation, despite changes in peak intensities, the OPLS-DA model continued to distinguish the EOs into six separate clusters (Fig. 2f), with R2X = 0.987, R2Y = 0.967, and Q² = 0.958. However, when a new sample spectrum is introduced, the method requires re-fitting. It cannot directly yield a predictive result, highlighting the need for the development of more advanced analytical approaches.

Fig. 2.

Raman spectral and cluster analysis of six citrus EOs before and after 28 days of volatilization. (a, b) Average Raman spectra of raw and volatilized EOs; (c, d) Deconvolution Raman spectra of raw and volatilized EOs; (e, f) OPLS-DA score plot of raw and volatilized EOs.

Since the Raman spectra of EOs are derived from their components, it is essential to identify the key bands that were used to distinguish different types of citrus EOs(Vargas Jentzsch et al., 2017). Raman spectra of citrus EOs exhibit well-known bands that can be readily identified. The strongest bands in the Raman spectra of Buddha's hand, lemon, lime, and sweet orange appear at 758, 1436, 1646, and 1678 cm⁻¹, indicating similar spectral profiles. In contrast, grapefruit and mandarin orange exhibited both shared and unique features. Grapefruit displayed similar strong bands at 756, 1436, and 1654 cm⁻¹, along with less intense peaks at 1262, 1300, and 1746 cm⁻¹. Meanwhile, mandarin orange exhibited similar strong bands at 760, 1436, 1638, and 1678 cm⁻¹, as well as less intense peaks at 1216, 1294, and 1332 cm⁻¹. For all the characteristic peaks in the RS of the six EOs, please refer to Table 1 for details. The compounds present in citrus EOs possess distinct functional groups that are closely related to their Raman spectral features. The main categories include terpenes (characterized by C=C stretching vibrations), oxygenated compounds such as alcohols, ketones, oxides, and aldehydes (exhibiting C-O/O-H stretching), aromatic compounds (featuring ring breathing modes), as well as various compounds containing CH3/CH2 bending modes. These molecular structures contribute significantly to the overall Raman spectral features. Specifically, the strong bands at 1646 and 1678 cm⁻¹ are attributed to C=C stretching, which is associated with the high terpene content in citrus EOs. The region from 1500 to 1300 cm⁻¹ is associated with methyl and isopropyl bending vibrations, with peaks between 1300 and 1436 cm⁻¹. In particular, the strong band at 1436 cm⁻¹ corresponds to CH3/CH2 bending vibrations. The presence of aromatic compounds in citrus EOs is evidenced by the band at 760 cm⁻¹, which is associated with ring breathing modes or ring deformation. Additionally, several low-intensity characteristic bands are observed in the region from 1300 to 1000 cm⁻¹, which may be related to the presence of oxygenated compounds exhibiting C-O/O-H stretching vibration in this spectral region. The absence of Raman peaks and/or peak shifts suggests that the spectra may result from the interaction of multiple molecules, with intermolecular interactions also contributing to the observed features (An et al., 2016).

Table 1.

Band assignments of characteristic peaks in the Raman spectra of six EOs at day 0 and day 28.

Wavenumber (cm-1)	Band Assignment	Buddha's Hand	Grapefruit	Lemon	Lime	Mandarin Orange	Sweet Orange	Ref.
Day 0
756	breathing mode (ring deformation)	–	+	–	–	–	–	Gloerfelt-Tarp et al. (2022)
758	breathing mode (ring deformation)	+	–	+	+	–	+	Gloerfelt-Tarp et al. (2022)
795	breathing mode (ring deformation)	+	–	+	+	+	+	Gloerfelt-Tarp et al. (2022)
760	breathing mode (ring deformation)	–	–	–	–	+	–	Gloerfelt-Tarp et al. (2022)
886	CH2 wagging	+	–	+	+	+	+	An et al. (2016)
1017	C−O stretching	+	–	+	+	–	+	Vargas Jentzsch et al. (2017)
1079	C−O stretching	+	+	+	+	–	+	Siatis et al. (2005)
1156	Conjugated O-H and C-H twisting	+	–	+	+	–	+	Lebanov and Paull (2021)
1216	C−O stretching	–	–	–	–	+	–	Lafhal et al. (2015)
1262	C-O-C vibrations	–	+	–	–	–	–	Gloerfelt-Tarp et al. (2022)
1294	Ether Ar-O- stretch (Ar = benzene ring)	–	–	–	–	+	–	Tyagi et al. (2022)
1300	= CH rocking vibrations	–	+	–	–	–	–	Lebanov and Paull (2021)
1332	CH3 bend (attached to a C=C)	–	–	–	–	+	–	ÇEbİ (2021)
1370	CH3 bend (attached to a C=C)	+	–	+	+	+	+	Lebanov and Paull (2021)
1436	CH3/CH2 bend	+	+	+	+	+	+	Jentzsch et al. (2015)
1638	conjugated C=C stretching vibration	–	–	–	–	+	–	Huang et al. (2022)
1646	conjugated C=C stretching vibration	+	+	+	+	–	+	Huang et al. (2022)
1654	conjugated C=C stretching vibration	–	+	–	–	–	–	Huang et al. (2022)
1678	nonconjugated C=C stretching	+	–	+	+	+	+	Huang et al. (2022)
1746	carbonyl stretching (C=O)	–	+	–	–	–	–	Tyagi et al. (2022)
Day 28
756	breathing mode (ring deformation)	–	+	–	–	–	–	Gloerfelt-Tarp et al. (2022)
758	breathing mode (ring deformation)	+	–	+	+	+	+	Gloerfelt-Tarp et al. (2022)
795	breathing mode (ring deformation)	+	–	+	+	+	+	Gloerfelt-Tarp et al. (2022)
886	CH2 wagging	+	–	+	+	+	+	An et al. (2016)
1017	C−O stretching	+	–	+	+	+	+	Vargas Jentzsch et al. (2017)
1079	C−O stretching	+	+	+	+	–	+	Siatis et al. (2005)
1156	Conjugated O-H and C-H twisting	+	–	+	+	+	+	Lebanov and Paull (2021)
1214	C−O stretching	–	–	–	–	+	–	Lafhal et al. (2015)
1260	C-O-C vibrations	–	+	–	–	–	–	Gloerfelt-Tarp et al. (2022)
1292	Ether Ar-O- stretch (Ar = benzene ring)	–	–	–	–	+	–	Tyagi et al. (2022)
1298	= CH rocking vibrations	–	+	–	–	–	–	Lebanov and Paull (2021)
1370	CH3 bend (attached to a C=C)	+	+	+	+	+	+	ÇEbİ (2021)
1436	CH3/CH2 bend	+	+	+	+	+	+	Lebanov and Paull (2021)
1638	conjugated C=C stretching vibration	–	–	–	–	+	–	Jentzsch et al. (2015)
1646	conjugated C=C stretching vibration	+	–	+	+	–	+	Jentzsch et al. (2015)
1652	conjugated C=C stretching vibration	–	+	–	–	–	–	Jentzsch et al. (2015)
1678	nonconjugated C=C stretching	+	–	+	+	+	+	Huang et al. (2022)
1744	carbonyl stretching (C=O)	–	+	–	–	–	–	Tyagi et al. (2022)

Note: “+” indicates the presence of a peak and “-” indicates its absence.

3.3. Identification model of citrus EOs at day 0 and day 28

This study applied ML algorithms to analyze the Raman spectra of citrus EOs to identify different citrus EO types. Five ML algorithms were developed and compared, including RF, DT, SVM, AdaBoost, and XGBoost. The performance of each model for day 0 and day 28 was comprehensively evaluated using a validation dataset and a test dataset via five evaluation metrics (Table 2). The results showed that both the SVM and RF algorithms achieved high accuracy (accuracy = 100 %) and stability (5-fold CV = 100 % on day 0, 5-fold CV = 99.70 % on day 28) in the internal validation sets. The remaining algorithms also yielded accuracies above 95 % in the internal validation sets. In the external test set, SVM maintained robust classification performance (accuracy = 100 %, 5-fold CV = 100 %), whereas the performance of all other models declined. Therefore, SVM represents an effective approach for distinguishing volatile EOs.

Table 2.

Performance comparison of five ML algorithms on the internal validation and external test sets to identify six citrus EOs at day 0 and day 28.

Algorithm	Internal Validation Set					External Test Set
	Accuracy	Precision	Recall	F1-score	5-Fold CV	Accuracy	Precision	Recall	F1-score
Day 0
RF	100 %	100 %	100 %	100 %	100.00 %	99.17 %	99.17 %	99.17 %	99.17 %
DT	100 %	100 %	100 %	100 %	98.66 %	97.50 %	97.50 %	97.50 %	97.49 %
∗SVM	100 %	100 %	100 %	100 %	100.00 %	100 %	100 %	100 %	100 %
AdaBoost	100 %	100 %	100 %	100 %	99.73 %	99.17 %	99.17 %	99.17 %	99.17 %
XGBoost	98.94 %	98.94 %	99.24 %	98.95 %	98.66 %	86.67 %	86.67 %	86.67 %	85.39 %
Day 28
RF	100 %	100 %	100 %	100 %	99.70 %	98.81 %	98.81 %	98.48 %	98.80 %
DT	96.43 %	96.43 %	95.99 %	96.42 %	98.21 %	95.27 %	95.27 %	95.27 %	95.30 %
∗SVM	100 %	100 %	100 %	100 %	99.70 %	100 %	100 %	100 %	100 %
AdaBoost	100 %	100 %	100 %	100 %	83.63 %	99.17 %	99.17 %	99.17 %	99.17 %
XGBoost	98.81 %	98.81 %	98.48 %	98.81 %	97.92 %	96.67 %	96.67 %	96.67 %	96.67 %

Note: RF: Random Forest; DT: Decision Tree algorithm; SVM: Support Vector Machine; AdaBoost: Adaptive Boosting; XGBoost: eXtreme Gradient Boosting; 5-Fold CV: 5-Fold Cross Validation. ∗The machine learning with the best predictive performance.

To improve the interpretability of the SVM model, we employed the class-specific weight vectors coef to examine the distribution of classification weights across different categories of Raman spectra at day 0 and day 28. The corresponding feature importance diagrams are presented in Supplementary Fig. S2. The 1600-1700 cm⁻¹ band is the most crucial feature, corresponding to the 1638-1678 cm⁻¹ peak generated by C=C stretching of conjugated and non-conjugated double bonds (An et al., 2016). The 1400-1500 cm⁻¹ band is the second most important feature, corresponding to the 1434-1436 cm⁻¹ peak associated with methyl and isopropyl bending vibrations (Vargas Jentzsch et al., 2021). The 740-810 cm⁻¹ band is the third most crucial feature, with peaks observed at 754-760 cm⁻¹ related to ring breathing modes and non-planar C-H vibrations (Siatis et al., 2005). After 28-day evaporation, the main spectral features are located at 1600-1700 cm⁻¹, 750-820 cm⁻¹, and 450-550 cm⁻¹. The 1600-1700 cm⁻¹ and 750-820 cm⁻¹ bands are consistent with those observed before evaporation. The 450-550 cm⁻¹ band contains several low-intensity peaks, related to Ring deformation (An et al., 2016). These characteristic peaks align with those observed in previous Raman spectra, indicating their importance in distinguishing between different citrus EOs. The C=C stretching vibration corresponds to compounds containing C=C stretching, such as terpenes. Methyl and isopropyl bending vibrations are typical of compounds with methyl or isopropyl substitutions. Ring breathing modes are observed in compounds with rings, such as aromatic compounds (Siatis et al., 2005). Differences in the relative contents of these compounds are key factors in GC-MS analysis that distinguish different EO categories.

3.4. Analysis of characteristic peaks in Raman spectra of citrus EOs

The analysis of characteristic peaks in RS across five time points focused on the typical peaks of the six citrus EOs identified in the study above (Fig. 3). The top panels magnify these spectral regions. At the same time, the violin plots on the right show intensity distributions at specific Raman shifts, with statistically significant differences marked by alphabetical letters. The abundance of Raman spectral features changed from day 0 to day 28. The peaks at the 1646 cm⁻¹ region, associated with C=C stretching, showed significant increases in intensity in all samples. The peaks at 1436 cm⁻¹, related to conjugated CH3/CH2 bending vibrations, showed substantial increases in intensity in all samples except grapefruit EO. Specifically, lemon, buddha's hand, and mandarin orange EO showed a clear upward trend, confirming their higher volatility and susceptibility to degradation. By comparison, grapefruit EO showed relatively mild changes, reflecting its higher chemical stability. Mandarin orange EO displayed significant spectral shifts at 1434 cm⁻¹ and 1646 cm⁻¹, associated with a decrease in terpenes and an increase in aldehydes and ketones from day 0 to day 28. In contrast, the grapefruit EO exhibited only minor spectral changes, which corresponded to relatively small variations in its VOC composition. The observed spectral differences result from chemical transformations in compounds containing conjugated systems, alkyl side chains, and ring structures. These transformations affect Raman-active vibrational modes and can be quantitatively captured through statistical analysis.

Fig. 3.

Statistical differences in Raman spectral peak intensities of six citrus EOs at five volatilization time points. (a) Buddha's Hand; (b) Grapefruit; (c) Lemon; (d) Lime; (e) Mandarin orange; (f) Sweet orange.

3.5. Predictive analysis of citrus EOs between 0 and 28 days

Classification of six citrus EOs across five time points, including the performance of the OPLS-DA model, classification accuracy of the algorithm, SVM regression curves, and interpretability analysis (Fig. 4). The OPLS-DA model performed supervised classification based on Raman spectral data, successfully separating samples into five distinct clusters. R2X values ranged from 0.914 to 0.969, R2Y values from 0.715 to 0.868, and Q² values from 0.688 to 0.852. Among all models, sweet orange achieved the best performance, with the highest R2Y (0.868) and Q² (0.852), indicating strong explanatory and predictive abilities. In contrast, lemon EO has the highest R2X (0.968) but the lowest R2Y (0.715) and Q² (0.688), reflecting limited predictive power. To accurately classify six citrus EOs at five time points, this study compared five ML algorithms to determine the optimal decision-making model. The performance of each model was comprehensively evaluated using a validation dataset and test dataset, and five evaluation metrics (Supplementary Tables S14–S19). Results showed that the SVM algorithm achieved high accuracy and stability. SVM regression analysis was employed to predict the evaporation duration of each EO and compared with the actual values. The results demonstrated that the SVM model achieved an R-squared for calibration ($R_C^2$), greater than 0.98, and a Root Mean Square Error for calibration (RMSEc) less than 0.12 for each EO dataset, indicating excellent model fitting and high accuracy.

Fig. 4.

Classification of citrus EOs at five volatilization time points. (a) Buddha's Hand; (b) Grapefruit; (c) Lemon; (d) Lime; (e) Mandarin orange; (f) Sweet orange. Each panel includes: (i) OPLS-DA score plot; (ii) Accuracy of five ML algorithms; (iii) SVM regression curve; (iv) Interpretability analysis.

The interpretability analysis of the SVM algorithm revealed three consistently high-weighted spectral regions: 1600-1700 cm⁻¹, 1300-1500 cm⁻¹, and 750-850 cm⁻¹. As mentioned earlier, these regions correspond to specific molecular vibrations and compound categories that are highly susceptible to oxidation, cyclization, and evaporation during evaporation. These processes lead to observable spectral shifts, which serve as important features for identifying and distinguishing EO categories. Similarly, our previous analysis of characteristic peaks showed significant temporal changes in spectral intensity within these high-weighted regions. Lemon EO exhibited more high-weight features, which were consistent with the upward trend observed in the Raman spectral peak intensities. Grapefruit EO displayed low-weight features and relatively mild changes in Raman peak intensity. The spectral feature distribution of other EOs was aligned with their key Raman peaks, confirming the importance of these regions in capturing evaporation-related changes.

4. Discussion

Citrus EOs are widely used in various industries due to their aromatic and biological properties. However, their volatile components are highly susceptible to degradation when exposed to environmental factors such as heat, light, and oxygen (Turek and Stintzing, 2012, 2013). This instability can compromise the quality and therapeutic efficacy of EOs, creating challenges for quality control and market regulation. Conventional methods for analyzing EOs, such as GC-MS, rely on specialized personnel and expensive equipment (Do et al., 2015; Epping and Koch, 2023). Therefore, there is an increasing need to develop a rapid and on-site detection method to ensure the authenticity and stability of citrus EOs during production, storage, and distribution.

In addition to conventional chromatographic techniques, spectroscopic methods have been applied for the authentication and analysis of EOs, as they provide detailed molecular-level information and reveal unique spectral fingerprints (Tyagi et al., 2022). However, spectroscopy is ineffective when comparing similar spectra or when the differences are very subtle (Lebanov and Paull, 2021). For instance, one study successfully detected the adulteration of turmeric EO with vegetable oil using Fourier-transform infrared (FTIR) spectroscopy combined with partial least squares regression (Cobbinah et al., 2024). Although effective, FTIR requires sample preparation, which limits its suitability for on-site detection. In contrast, RS is more cost-effective due to its preparation-free nature (Vargas Jentzsch et al., 2017). Another study demonstrated that a smartphone-based handheld Raman spectrometer, combined with RF and partial least squares discriminant analysis, could accurately identify and quantify adulteration in ylang-ylang EO (Lebanov and Paull, 2021). These studies demonstrate that integrating chemometrics and supervised learning enables accurate sample classification and prediction by uncovering subtle spectral differences (Cebi et al., 2021; Tang et al., 2025), providing a promising direction for EO detection.

Therefore, in this pilot study, we combined RS with ML technology to analyze Raman spectra of six citrus EOs at five time points, aiming to evaluate the performance of the technique in identifying closely related EOs and assessing their subtle changes during volatilization for quality control. Our analysis revealed that the SVM model can accurately distinguish EO samples before and after evaporation, achieving 100 % accuracy in the external test sets, while consistently tracking changes in EO quality over time. Meanwhile, GC-MS analysis was performed before and after evaporation to assess compositional changes. The observed changes reflected the loss of volatile components (Fig. 1), supporting the use of RS for distinguishing EOs at different stages of evaporation and providing a basis for developing an RS-ML detection method.

However, due to the complex chemical composition of citrus EOs, it remains challenging for RS to identify their components accurately. To address this limitation, we applied spectral deconvolution based on molecular vibrational features, which has been successfully used to refine and distinguish differences between spectra (Li et al., 2024). The Raman spectra of the six EOs in this study are composed of multiple common and unique characteristic peaks (Table 1), and the structural differences among these peaks may serve as spectral markers for sample identification (Si et al., 2024). The distinctive peaks at 1646 and 1678 cm⁻¹ correspond to C=C stretching vibrations of conjugated and non-conjugated double bonds, typically arising from terpenes (An et al., 2016). The peak at 1436 cm⁻¹ is attributed to CH3/CH2 bending vibrations from various alkyl-containing compounds (Vargas Jentzsch et al., 2021). The peak at 758 cm⁻¹ corresponds to ring breathing modes associated with aromatic compound structures (Siatis et al., 2005). The results indicate that characteristic Raman peaks differ among six citrus EOs and change after evaporation, demonstrating that RS can effectively distinguish EO samples. The supervised algorithm OPLS-DA was used to explore and model the clustering of different EO samples, demonstrating effective group separation with strong explanatory and predictive power (Fig. 2, Fig. 4). However, as a software-based method, it requires reparameterization and refitting for each new dataset, unlike ML models that support direct prediction. To address this, we leveraged the clustering patterns among EO samples to develop a robust ML model for classifying Raman spectra.

There are many methods to apply RS-ML algorithms to EO detection, such as EO classification and the identification of adulterants in pure EOs (Zhao et al., 2022). However, while most existing studies have focused on sample identification or detection, our work developed both an EO classification model and a model for quality assessment. In this study, we constructed five ML models (RF, DT, SVM, AdaBoost, and XGBoost), optimized their hyperparameters via grid search, and used various evaluation metrics to assess model performance. The results show that the SVM model performs well in distinguishing EOs before and after evaporation, as well as across five evaporation time points (Table 2 and Supplementary Table S16-21S14–19). While ML has achieved remarkable performance in RS studies of EOs, the models are often seen as “black boxes,” potentially leading to inaccurate or undesired outcomes (Tang et al., 2024). Interpretability methods such as Shapley Additive exPlanations have been successfully used in other studies to identify key spectral features in ML models (Yuan et al., 2025). Similarly, we employed the class-specific weight vectors coef to reveal important Raman bands, enhancing model transparency (Supplementary Figure S2 and Fig. 4). The results show that the model was able to specifically identify Raman characteristic peaks unique to different EO samples, primarily located in three regions: 1600-1700 cm⁻¹, 1300-1500 cm⁻¹, and 750-850 cm⁻¹. These peaks corresponded to characteristic Raman signals and were correlated with the compositional differences of the EOs observed in the GC-MS analysis, indicating that the model captures compositional alterations of EO and thereby enhances its credibility in quality assessment.

Nevertheless, this study has certain limitations and unresolved challenges that remain to be addressed. First, the model was trained on a limited set of citrus EO types under controlled conditions, which may not fully capture the diversity found in commercial products. Second, variations in instrument models, wavelengths, and detection ranges can affect the consistency of Raman spectra, rendering the method developed in this study inapplicable across platforms. In the future, we should aim to establish cross-platform multimodal models to enable the integration of heterogeneous resources. Thirdly, batch-to-batch variability caused by factors such as the origin or manufacturer of EOs may alter their compositional profiles, underscoring the need to expand the sample size further.

In summary, the RS-ML method developed in this study enables the rapid, accurate, and cost-effective identification and quality assessment of citrus EOs. This approach complements existing GC-MS techniques by offering on-site traceability, authenticity verification, and quality monitoring. With continued refinement, this method holds significant potential for use in industrial quality control and regulatory applications.

5. Conclusion

In this study, we developed a rapid and on-site detection method for citrus EOs by integrating RS with ML algorithms. GC-MS analysis validated the authenticity of the citrus EOs and revealed differences in their composition after evaporation. By examining both the average Raman spectra and deconvoluted spectral features of six citrus EOs, we found that RS could capture subtle differences between EO samples, which was further supported by clustering results from OPLS-DA. Using an SVM model, the method effectively distinguished the Raman spectra of citrus EOs across different species and five evaporation time points, achieving classification accuracies ranging from 92.96 % to 100 %. In summary, the integration of RS and SVM offers a rapid, cost-effective, and reliable on-site solution for citrus EO analysis, providing strong potential for efficient identification and quality assessment, which is crucial for quality control and market regulation.

Data availability statement

The datasets are available from the corresponding author on reasonable request.

Author contributions

LW and QHL conceived and designed the experiments. LW provided platforms and resources and contributed to project administration and student supervision. LW contributed to the funding acquisitions. YXH, JWT, JC, YYX, ZWM, and QHL carried out the computational and experimental investigations. All authors contributed to the writing of the manuscript and approved the submission of the manuscript.

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Handling Editor: Professor Georgios Leontidis

Appendix A. Supplementary data

The following are the Supplementary data to this article: